HeyHomees Lab Notes · No. 11 · Liquid AI · Published October 28, 2026
By HeyHomees Newsroom ·
In 30 seconds
- A 1.73B-parameter text model for local agents.
- Runs in about 1 GB memory at 4-bit precision, fitting phones and laptops.
- Supports a 256K token context window and tool-calling capabilities.
What is it?
LFM3-1.7B is a text-only language model in the Liquid Foundation Model family, intended for local agents, tool calling, structured outputs, and domain-specific fine-tuning.
How does it work?
It uses a tri-hybrid layer mix of 13 gated short convolutions, 5 sliding-window attention layers, and 6 global attention layers across 24 layers; the model was post-trained on agentic traces and tool calls and can run at 4-bit precision using about 1 GB of memory.
How well does it do?
On the Claw-Eval General (EN) benchmark it scores an average of 68.64, and on the ToolSandbox benchmark it reaches 80.36; for coding it attains 73.46 on LiveCodeBench v6, and for instruction following it gets 96.66 on IFEval.
What does it need to run?
At 4-bit precision the model fits in roughly 1 GB of memory, allowing inference on phones, laptops, or any GPU-enabled device; supported inference frameworks include vLLM, SGLang, llama.cpp, MLX, and ONNX.
By the numbers
- Parameters: 1.73B
- Context length: 256K tokens
- Memory footprint at 4-bit: about 1 GB
- Architecture layers: 24 layers: 13 gated short convolutions, 5 sliding-window attention, 6 global attention
- Vocabulary size: 128,000 tokens
- Claw-Eval General (EN) average: 68.64
- ToolSandbox score: 80.36
- LiveCodeBench v6 score: 73.46
The honest read
The reported benchmark numbers come from the model card and reflect the authors' own evaluations; they do not include comparisons to other labs' models beyond the table shown. Scores such as the 68.64 on Claw-Eval and 80.36 on ToolSandbox demonstrate capability on specific agentic and tool-use tasks, but the model card notes it is not recommended for complex multi-step reasoning or agentic coding due to its small size. The memory claim of "about 1 GB" applies only when the model is quantized to 4-bit precision; larger precisions would require more RAM. Performance figures are limited to the listed benchmarks and do not cover latency, power consumption, or real-world user experience, so they give an incomplete picture of deployment suitability.
What we don't know yet
- Exact inference latency on different hardware.
- Training data composition beyond token counts.
- Power or energy consumption during inference.
Sources
- Model card, Hugging Face (primary)
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.